Classifications

• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T1/00—General purpose image data processing
  • G06T1/60—Memory management
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F12/00—Accessing, addressing or allocating within memory systems or architectures
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F12/00—Accessing, addressing or allocating within memory systems or architectures
  • G06F12/02—Addressing or allocation; Relocation
  • G06F12/08—Addressing or allocation; Relocation in hierarchically structured memory systems, e.g. virtual memory systems
  • G06F12/0802—Addressing of a memory level in which the access to the desired data or data block requires associative addressing means, e.g. caches
  • G06F12/0862—Addressing of a memory level in which the access to the desired data or data block requires associative addressing means, e.g. caches with prefetch
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F12/00—Accessing, addressing or allocating within memory systems or architectures
  • G06F12/02—Addressing or allocation; Relocation
  • G06F12/08—Addressing or allocation; Relocation in hierarchically structured memory systems, e.g. virtual memory systems
  • G06F12/0802—Addressing of a memory level in which the access to the desired data or data block requires associative addressing means, e.g. caches
  • G06F12/0893—Caches characterised by their organisation or structure
  • G06F12/0897—Caches characterised by their organisation or structure with two or more cache hierarchy levels
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F12/00—Accessing, addressing or allocating within memory systems or architectures
  • G06F12/02—Addressing or allocation; Relocation
  • G06F12/08—Addressing or allocation; Relocation in hierarchically structured memory systems, e.g. virtual memory systems
  • G06F12/0802—Addressing of a memory level in which the access to the desired data or data block requires associative addressing means, e.g. caches
  • G06F12/0806—Multiuser, multiprocessor or multiprocessing cache systems
  • G06F12/0815—Cache consistency protocols
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F12/00—Accessing, addressing or allocating within memory systems or architectures
  • G06F12/02—Addressing or allocation; Relocation
  • G06F12/08—Addressing or allocation; Relocation in hierarchically structured memory systems, e.g. virtual memory systems
  • G06F12/0802—Addressing of a memory level in which the access to the desired data or data block requires associative addressing means, e.g. caches
  • G06F12/0875—Addressing of a memory level in which the access to the desired data or data block requires associative addressing means, e.g. caches with dedicated cache, e.g. instruction or stack
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F12/00—Accessing, addressing or allocating within memory systems or architectures
  • G06F12/02—Addressing or allocation; Relocation
  • G06F12/08—Addressing or allocation; Relocation in hierarchically structured memory systems, e.g. virtual memory systems
  • G06F12/0802—Addressing of a memory level in which the access to the desired data or data block requires associative addressing means, e.g. caches
  • G06F12/0877—Cache access modes
  • G06F12/0879—Burst mode

Definitions

• the present invention relates to cache memory structure and management in digital data processing, and particularly, in digital image data processing.
• a cache memory is generally a shallow and wide storage device, inside or close to a processor that facilitates processor's access to the data and content change of the data.
• cache memory management The philosophy of cache memory management is to retain copies of data and instructions which are often used, or are most likely to be used in near future by the processor, inside the fastest accessible storage device. This makes the access of a processor to data and instructions many times faster than to otherwise access them in an external memory. However, care must be taken in such operations as changing content in cache memory and in external memory should be harmonized. These issues, with their hardware and software features, have created the art of cache memory structure and management. As mentioned, a cache memory keeps copies of data and address pointers that are most likely to be accessed next by the processor. An external memory typically holds data in capacitors and needs refresh cycles to replenish the charge on the capacitors to prevent the loss of data.
• U.S. patent No. 6,681 ,296 to Liao presents a microprocessor with a control unit and a cache, which is selectively configurable as single or partitioned with locked and normal portions.
• U.S. patent No. 6,721,856 to Arimilli presents a cache with coherency state and system controller information of each line with different subentries for different processors containing a processor access sequence.
• US patent 6,629,188 discloses a cache memory with a first and a second plurality of storage spaces.
• U.S. patent No. 6,295,582 discloses a cache system with data coherency and avoiding deadlock with substantially sequential read and write commands.
• U.S. patent No. 6,339,428 discloses a cache apparatus in video graphics where compressed texture information are received and decompressed for texture operations.
• U.S. patent No. 6,353,438 discloses a cache organization with multiple tiles of texture image data and directly mapping of data into cache.
• the present invention in one aspect provides a method for cache memory management and structure in digital data processing, and in particular, in digital image processing in a setting consisting of:
• the method uses the following cache structure: (i) a deeper secondary cache memory (SCM) with higher storage capacity, having a plurality of banks and each bank having a plurality of storage lines, to read data from said external memory;
• SCM secondary cache memory
• a faster and wider primary cache memory with lower storage capacity, having a plurality of banks and each bank having a plurality of storage lines, from where data are read by said PU2; and
• a control logic containing control stages and control queues, providing pre-fetching and cache coherency; to access data in the external memory, upon receiving address sequences and control parameters form the PU1 , and to prepare data for fast access and processing by the PU2.
• the method achieves cache coherency and hides memory read latency via:
• the present invention provides a cache system based on the method just described. Further details of different aspects and advantages of the embodiments of the invention will be revealed in the following description along with the accompanying drawings.
• FIG. 1 represents an overall scheme of a cache system built in accordance with the present invention
• FIG. 2 represents the detailed structure of a cache system built in accordance with the present invention
• FIG. 3 represents an example of the block structure of the input data to be cached
• FIG. 4 represents the general structure of the primary cache system built according to the present invention
• FIG. 5 represents the general structure of the secondary cache system built according to the present invention.
• FIG. 6 represents the flow logic of a cache system built according to the present invention.
• the present invention relates to cache structure and management.
• the implemented example, given in this description, is that of image processing with simultaneous coordinate transformation.
• a plurality of processors attempt to fetch data and control parameters from an external memory and other processors with an arbitrary format.
• two-dimensional (2D) image transformation example, given here could be trivially replaced by any 2D data transformation without departing from the scope of the present invention.
• data image pixel data.
• the plurality of processors which issue control parameters regarding the structure and topology of input data as the geometry engine.
• the plurality of processors accessing data for operation as filter engine, and the corresponding operations as filtering.
• FIG. 1 is an illustrated example of the setting of cache system 100 in a computing arrangement, designed for digital image data processing, with simultaneous coordinate transformation.
• Cache system 100 interfaces with two sets of processors.
• the first plurality of processors in this example implementation, constitute geometry engine 300 and the second plurality of processors constitute filter engine 500.
• cache system 100 interfaces with external memory 700, which could be any memory with access latency.
• Cache system 100 receives control parameters including coordinate transformation as well as filter footprint parameters from geometry engine 300. It simultaneously receives pixel data from external memory 700.
• Cache system 100 provides these data to the filter engine 500 in a manner as to optimize the filtering process with minimal stalling of filter engine 500.
• each output pixel is formed based on information from many input pixels.
• the output pixel coordinates are mapped onto input pixel coordinates; This is a coordinate transformation, normally achieved electronically via image warping techniques.
• a filtering or sampling function is needed to generate output pixel specifications, namely, intensities of the constituent colors, and other information such as sampling format and blending function.
• the area containing all the pixels around the central input pixel, over which the sampling is performed, is called the filter footprint. It is well known in the art that the size and shape of the filter footprint affect the quality of the output image.
• cache system 100 The function of cache system 100 is to use a dedicated architecture and pre-fetching logic to provide enough random access pixel data and control parameters to filter engine 500 so it has data to process at any given clock rate with minimal stalling. With an optimally sized read request queue, cache system 100 is able to hide the majority of the memory read latency inherent in external memory 700, from where the pixel data are fetched. This hiding of the memory read latency is paramount to the filter performance. If the latency is not hidden properly, filter engine 500 will not have maximized throughput. The amount of allowable stalling is a design parameters. One needs to adjust different parameters to achieve required throughput as a tradeoff with hardware cost. In addition, cache system 100 provides a control path for the coordinate transformation and filter footprint parameters, read from geometry engine 300. Cache system 100 ensures that the pixel data from external memory 700 on the one hand, and control parameters from geometry engine 300 on the other hand, are synchronized when they arrive at the input to filter engine 500.
• FIG. 2 is an illustrated example of the detailed structure of cache system 100.
• cache system 100 receives certain control parameters from geometry engine 300. These parameters include the coordinates of the mapped input pixel, U and V, and additional control parameters, including those defining the shape, rotation, and size of the filter footprint.
• cache system 100 receives pixel data for each of the pixels included in the filter footprint from external memory 700. These data include the intensity levels of constituent colors in color space, e.g. RGB or YCrCb, sampling format, e.g. 4:4:4 or 4:2:2, and blending function, i.e. with ⁇ or without ⁇ .
• the structure of cache system 100 is related to dividing the input image into blocks of size m x n pixels.
• Input image 330 comprises a certain number of pixels, for instance, 1024 x 1024, grouped into blocks.
• Each input pixel block 332 contains m x n input pixels 334.
• the structure of blocks is in general a function of footprint shape and size in different filtering schemes.
• Cache system 100 fetches data relating to m x n input pixel blocks 332 and generates data blocks usable by filter engine 500. As such, the system has to determine which blocks fall inside the footprint and which pixels inside these blocks must be included for filtering.
• the structure of cache system 100 is scalable to match the input block data structure. It should also be noted that in general the structure of cache system 100 is a function of the nature and structure of the operation of filter engine 500. In the particular case of image processing, the structure and topology of the operation are defined partially by the filter footprint.
• cache system 100 comprises primary cache 110 which is shallow and wide with lower capacity, secondary cache 120 which is deep with higher capacity, block inclusion stage 150, block data generation stage 130, primary cache control stage 170, and secondary cache control stage 190.
• the pixel data are first read into secondary cache 120 from external memory 700. Then these data are reformatted and uncompressed by block generation stage 130 for use by filter engine 500. These reformatted data are put into a queue to be placed in primary cache 110 at the appropriate time, where they are readily accessible by filter engine 500. Below we explain the data path and the control logic structure respectively.
• secondary cache 120 is a higher capacity storage device that reads raw data from external memory 700.
• the pixel data in external memory 700 are stored in an arbitrary format, generally not well suited for processing in filter engine 500, for instance, in a particular example, the data are stored sequentially, in scan- line order.
• Secondary cache 120 is designed to read these data efficiently with minimal interruption.
• Each line in the secondary cache is designed to accommodate a burst of b 2 bytes of data from external memory 700. For this reason, each line in secondary cache 120 is sized according to the structure of the external memory 700 and the read requirements.
• the number of lines in secondary cache 120, in which the data are stored, is also a design parameter optimized to reduce the secondary cache miss count.
• Secondary cache 120 is additionally banked to allow a read throughput sufficient to update primary cache 110 to minimize the stalling of filter engine 500.
• secondary cache 120 is designed to have a certain number of banks with independent access lines to read data from external memory 700 simultaneously.
• secondary cache 120 has a number of banks 122, each with a certain number of lines 124.
• Each secondary cache line contains data from one data burst read from external memory 700. These data need to be eventually read by filter engine 500.
• the number of secondary cache banks is designed as a function of the data throughput. For an m x n input block structure and a required number of clock cycles, N 0 , to read the data, n / N 0 banks are needed in secondary cache 120.
• the combination of U and V least significant bits (LSBs) is used.
• Block generation stage 130 reads data in secondary cache 120 and prepares these data in blocks that include all the data from an m x n input pixel block. As described above, block generation stage 130 reads from n /Nc lines of secondary cache 120 per clock cycle. This ensures that in each Nc clock cycles, all the data relating to one input pixel block are read simultaneously. Depending on the packing format of the data and throughput requirements, multiple reads may be required from secondary cache 120 to generate the input pixel block. In addition to reading these data, block generation stage 130 is adapted to reformat and uncompress these data into a format readily usable by filter engine 500.
• Block generation stage 130 therefore hides the original pixel data format, which could be compressed with various compression schemes. This exonerates filter engine 500 from figuring out the format of the pixel data in external memory 700 and unpacking the original formatted data into blocks that are usable for filtering. These block data are eventually stored in primary cache 110, from where they are read by filter engine 500.
• primary cache 110 is designed in a fashion to optimize the rate of data access in filter engine 500. As such, it has a shallow but wide structure for multiple lines of access.
• Primary cache 110 is divided into a certain number of banks, with each primary cache bank 112 being read independently and simultaneously by filter engine 500. The number of primary cache banks is determined according to empirical data and simulation to optimize filtering performance.
• Each primary cache bank 112 contains a certain number of primary cache lines.
• Each primary cache line 114 contains data from an entire m x n block of input data.
• filter engine 500 reads data containing hi input blocks per cycle in proper format. This is crucial since for sampling, many input blocks around an input pixel are needed and if they are not provided to filter engine 500, it will stall. The amount and frequency of stalling determine the throughput performance.
• each primary bank 112, inside primary cache 110 is also divided into a certain number of partitions. As explained above, a certain number of LSBs are used to distribute the data amongst different primary cache banks. In the remaining bits of the input pixel U and V addresses, further LSBs are used again to distribute data in each primary cache bank 112. For 2 f lines per primary cache bank and g LSBs used to partition each bank, this division yields a 2 f /2 9 set-associative architecture.
• This design is again used along with proper replacement policy for primary cache 110, which will be explained later, to achieve optimal throughput.
• This architecture is scalable in a simple and natural way since for larger input data volume more bits are available in the U and V addresses.
• FIG. 6 represents cache control logic 400. This logic structure controls reading data from external memory 700 by secondary cache 120, reading and reformatting data in block generation stage 130, and data block storage in primary cache 110.
• step 402 it is determined which blocks of data are necessary for sampling based on control parameters received from geometry engine 300.
• step 410 it is determined whether these data are present inside the primary cache. If present, an entry is written to primary control queue at step 412 and the address of these data is sent to the filter engine at step 414. If the data are not present in the primary cache, at step 415, according to an adopted replacement policy explained later, it is determined which primary cache line to replace. Then the address of this primary cache line is written to the primary control queue at step 416 and sent to the filter engine at step 418. It is then determined whether these data are present in the secondary cache at step 420. If the data are not present there either, it is decided at step 422 which secondary cache lines to replace. Then a read request is sent to the external memory to fetch the data that are later read into the secondary cache at step 426. If the data are present in the secondary cache, an entry is written into secondary cache control queue at step 428.
• secondary cache data are read for block generation at step 440.
• the data are read from multiple secondary cache banks and are reformatted and uncompressed at step 442.
• a block of input data in the proper format is sent into a queue to be stored in the primary cache. These data are stored in primary cache banks at step 452.
• the pre-fetching logic is designed to hide the read latency in filter engine 500. Without this control logic structure, data throughput will not be optimal and filter engine 500 will have a higher rate of stalling. With sufficiently sized queues, optimal storage sizes, data preparation, and intelligent replacement policy, cache system 100 hides most of the read latency by running ahead of filter engine 500.
• Block inclusion stage 150 is the starting point of the control logic. For each output pixel, it receives control parameters from geometry engine 300, including coordinates of the mapped input pixel and the shape of the filter footprint. Based upon the input pixel coordinates, U and V, the footprint shape, and other control parameters, the block inclusion logic determines which input blocks are required for processing each output pixel and which pixels in each block are required for sampling.
• Block inclusion stage 150 compares the coordinate positions of adjacent blocks with the geometry of the footprint to include blocks of pixels necessary for sampling.
• the block inclusion logic generates k blocks per clock cycle with each block differing in at least 1 U or 1 V least significant bit (LSB) in its block address. This guarantees that k combinations of LSB's will be present in each set of blocks generated by the block inclusion logic. This constraint is used to distribute the blocks amongst the primary cache banks.
• the number of generated blocks per cloc cycle, k is a function of the footprint size, and the topology of the blocks is a function of the footprint shape.
• Primary cache control stage 170 provides control logic for data handling in primary cache 110. For each input block determined by block inclusion stage 150, primary cache control 170 checks to see if the block is present in primary cache 110. If the data is present, this is termed a cache hit. Else a cache miss is registered and the miss flag is sent to secondary cache control 190. Primary cache control stage 170 writes an entry into primary control queue 212, indicating the address of the data inside the primary cache 110, as well as whether there has been a primary cache hit or miss. Primary control queue 212 is read by filter engine 500 on a FIFO basis. If a cache- miss flag is raised in one of the entries, filter engine 500 sends a read request to block queue 214 which will update primary cache 110.
• Secondary cache control stage 190 upon receiving a primary cache miss flag, will determine which steps to take to generate the m x n block that will be written into the primary cache. Secondary cache control stage 190 first determines whether the data exist in the secondary cache 120. This will yield a secondary cache hit or a secondary cache miss. If a secondary cache miss occurs, secondary cache control 190 sends a read request to external memory 700 to fetch the missing data into secondary cache 120 from external memory 700 and writes an entry into secondary control queue 216. If a secondary cache hit occurs, secondary cache control stage 190 does not send a read request and only writes an entry into secondary control queue 216, where entries are read by block generation stage 130 on a FIFO basis.
• block generation stage 130 Upon receiving each queue entry, block generation stage 130 reads raw data relating to an entire input block from secondary cache 120. These data are then reformatted in block generation stage 130 into a format readily usable by filter engine 500. Depending on data packing mode, multiple secondary cache lines maybe required to generate a primary cache line 114. After obtaining all the data relating to one input block and reformatting these data, block generation stage 130 writes an entry into block queue 214. Each block queue entry therefore contains all the data from the entire input block in proper format. Block queue entries are then received by primary cache 110, where they are stored to be readily accessed by filter engine 500. Accordingly, block queue 214 allows secondary cache 120 to run ahead of filter engine 500.
• cache system 100 is based on coherency of pixel data and control parameters in addition to the dedicated pre-fetching logic. No data are read by the secondary cache 120 without a request from secondary cache control stage 190. Once that data are in secondary cache, only entries in secondary control queue 216 determine whether these data are needed for block generation in block generation stage 130. Once a block of data is generated, they are put in a queue to be stored in primary cache 110 only upon a read request from filter engine 500, which is itself instigated by an entry in primary control queue 212. Moreover, filter engine waits for the arrival of both pixel data as well as control parameters from two independent queues before processing the data.
• the filter engine will process these data sequentially.
• the read latency was on the order of 128 clock cycles.
• the size of the queues in the present invention can be adjusted to match the memory read latency seen in the system and, as such, they are scalable design parameters based upon the system specifications.
• primary cache control stage 170 and secondary cache control stage 190 compare the central input pixel U and V coordinates with those of the existing block data in the cache lines. The entry with the largest distance from the central input pixel is then replaced. This policy stems from the fact that the closer the distance to the central pixel, the higher the probability of being needed for sampling calculations.
• the cache replacement policy is least-recently-used (LRU) based.
• primary cache control stage 170 and secondary cache control stage 190 opt to replace the cache lines that are least recently used.
• the design of cache system 100 has a few measures to make sure this system is scalable.
• the size of secondary cache lines is scalable to the memory read size, e.g.
• the number of secondary cache lines is scalable based on the required cache efficiency.
• the number of secondary cache banks is scalable based upon the input block data structure and the number of clock cycles per access out of the secondary cache.
• Scaling secondary cache 120 is based on size requirements and the cache system efficiency, i.e., the amount of input digital data to be reread.
• the number of blocks generated per clock cycle in block inclusion stage 150 is scalable based on filtering algorithm and footprint size and required throughput.
• the partitioning of primary cache 110 and secondary cache 120, based on the U and V input pixels LSBs is adaptable to the size of the cache. This is implemented by the number of bits used for a particular partitioning.
• the size of primary cache lines is scalable based on input block size.
• the number of primary cache banks is scalable based on filtering throughput.
• the sizes of different queues are also scalable parameters depending on memory latency vs. the required throughput. These sizes are determined based on simulations and empirical data.

Landscapes

• Engineering & Computer Science (AREA)
• Theoretical Computer Science (AREA)
• Physics & Mathematics (AREA)
• General Physics & Mathematics (AREA)
• General Engineering & Computer Science (AREA)
• Memory System Of A Hierarchy Structure (AREA)

Priority Applications (5)

Applications Claiming Priority (1)

Publications (1)

Family

ID=35907684

Family Applications (1)

Country Status (5)

Cited By (1)

Families Citing this family (8)

Citations (3)

Family Cites Families (7)

• 2004
  • 2004-07-14 JP JP2007521441A patent/JP5071977B2/ja not_active Expired - Fee Related
  • 2004-07-14 CN CNB2004800427711A patent/CN100533403C/zh not_active Expired - Fee Related
  • 2004-07-14 KR KR1020067023350A patent/KR101158949B1/ko active IP Right Grant
  • 2004-07-14 EP EP04757053A patent/EP1769360A4/de not_active Withdrawn
  • 2004-07-14 WO PCT/US2004/022878 patent/WO2006019374A1/en not_active Application Discontinuation

Patent Citations (3)

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events